27. Spread work across multiple GPUs
Program more than one GPU with peer-to-peer copies, NVLink, topology awareness, partitioning, and workload distribution. You will build a simple multi-GPU pipeline and measure whether scaling is actually happening.